AI Model Predicts Secondary-School English Performance With Built-In Uncertainty

A new computational framework claims the best mean predictive performance across four English learner corpora, forecasting writing scores and speaking-proficiency levels for secondary-school students more accurately than conventional statistical models, recurrent neural networks, and pretrained Transformer baselines. The framework, called HierarchicalFactorizedUncertainty, was developed by researchers at the University of Malaya and published in Frontiers in Molecular Neuroscience. Its authors are careful, however, to draw a sharp line between what the model actually does and what its title might suggest: neuroplasticity serves only as a theoretical lens, and no neural measurements were taken.

The study addresses a persistent problem in educational data science. Learner performance in secondary-school English is expressed through many observable dimensions, including written production, spoken performance, and proficiency assessments, and observations differ in task type, learner background, corpus size, and whether repeated measurements exist. Standard scoring models can represent learner language well, but they typically lack an explicit mechanism for jointly representing multi-level predictive structure, compact associations with recorded task or contextual variables, sparse interaction patterns, and distinct forms of predictive uncertainty within one unified system.

HierarchicalFactorizedUncertainty combines three jointly optimized components. A Hierarchical Composition Mapper learns structured, multi-level representations from learner-language evidence such as essays or transcribed speech. An l1 sparsity penalty then shrinks weak or redundant interaction factors toward zero, concentrating the model on a smaller subset of predictive relationships. Because the datasets are observational, these associations are treated as predictive dependencies, not causal effects of instructional manipulation.

The third component, an Uncertainty Propagation Resolver, separates two conceptually distinct sources of predictive uncertainty. Aleatoric uncertainty captures observation-dependent variability that remains even for a fixed model, including measurement noise and heterogeneity in outcomes; it is estimated from an observation-dependent predictive variance produced by a heteroscedastic head. Epistemic uncertainty reflects the limits of the learned model itself, arising from finite training evidence or incomplete coverage of the feature space; it is estimated using Monte Carlo dropout, with thirty stochastic forward passes at inference time and disagreement among the resulting predictive means quantifying model uncertainty. The total predictive uncertainty is their sum, and the final point prediction averages across the stochastic realizations.

The empirical evaluation draws on four complementary corpora. CSEE contains 13,270 publicly released essays from Chinese high-school examinations, with teacher-assigned overall writing scores on a nominal 0-to-20 scale as the target. ELLIPSE includes 6,482 essays by English language learners in Grades 8 through 12 in the United States, with holistic proficiency scores. SEEFLEX comprises 575 curriculum-based EFL examinations from upper-secondary students in Germany, spanning 1,979 individual texts and roughly 625,000 words. LOCSE supplies the longitudinal component: 104 Japanese senior-high-school learners assessed at eight occasions over 23 months, yielding 832 learner-occasion observations of transcribed monologic speaking, rated on an ordered proficiency scale. Texts were encoded with a fixed pretrained RoBERTa-base encoder, and analytic trait scores were excluded from predictive inputs to prevent label leakage.

Under controlled partitions and ten repeated training runs per configuration, the proposed model achieved the best mean value on every reported metric. On CSEE it reached a mean absolute error of 1.10 and root mean squared error of 1.49, with Spearman’s rank correlation of 0.847 and quadratic weighted kappa of 0.852, improving on the strongest baseline, DeBERTa-v3, by roughly 5.2 percent in MAE and 4.5 percent in RMSE. On ELLIPSE it obtained an MAE of 0.319 and RMSE of 0.445. On SEEFLEX, MAE fell from 5.78 to 5.41 relative to DeBERTa-v3, and on LOCSE the model achieved an MAE of 0.451 with quadratic weighted kappa of 0.751.

The statistical picture is more nuanced than the headline numbers. Pairwise two-sided Welch tests with Holm-Bonferroni correction showed that most differences from DeBERTa-v3 remained significant, but the advantage in Spearman’s rank correlation on CSEE did not survive multiplicity correction, dropping from a corrected significance level. The authors note that these tests characterize reproducibility of performance differences across stochastic training runs under a fixed test partition, not population-level sampling uncertainty. The framework’s advantage in absolute prediction error and score agreement on CSEE was therefore clearer than its advantage in rank ordering.

Ablation experiments, reported as patterns in mean performance without separate significance tests, distributed the credit across components. Removing the Hierarchical Composition Mapper produced the largest deterioration among architecture- and regularization-related ablations on all four datasets; on CSEE, MAE rose from 1.10 to 1.21, and on LOCSE it rose from 0.451 to 0.512. Removing the Factorized Relation Projector had a comparatively pronounced effect on SEEFLEX, where heterogeneous curriculum-based tasks make context modeling valuable. The two uncertainty mechanisms showed dataset-dependent contributions: on the writing corpora, removing aleatoric uncertainty hurt slightly more than removing epistemic uncertainty, while the ordering reversed on SEEFLEX and LOCSE, where disabling epistemic uncertainty raised LOCSE MAE from 0.451 to 0.501. Removing both was consistently worse than removing either alone, and retaining both yielded the strongest performance.

The authors frame the model as potential decision support for teachers rather than an automated judge. A low predicted score accompanied by low uncertainty could prioritize an essay for closer human review, whereas a similar prediction with high uncertainty should trigger collection of additional evidence, such as another writing sample or teacher-rated formative assessment, before any consequential decision. For longitudinal speaking data, sustained low predicted proficiency with low uncertainty might motivate focused practice, while abrupt predicted changes accompanied by high epistemic uncertainty should be verified with an additional speaking sample. These workflows are prospective: the study did not prospectively evaluate classroom deployment, and the model itself does not diagnose specific linguistic errors.

The interpretive boundaries are explicit throughout. The datasets contain no measurements of synaptic modification, neural reorganization, or any other biological marker of plasticity, so the learned latent representations are computational descriptions optimized for prediction, not neural states. Variation across the eight LOCSE occasions represents longitudinal change in observed speaking performance, not direct evidence of neural restructuring. Because all four datasets are observational secondary data, relationships involving contextual variables are predictive associations rather than causal effects of instruction. The implementation also uses a fixed hierarchical depth of three levels, may not transfer to other school subjects without independent validation, and omits socio-emotional factors such as motivation or language anxiety that the corpora do not observe. The authors call for prospective classroom studies and, for any future claims about neuroplastic change, for combining educational observations with neuroimaging or electrophysiological measurement rather than inferring neural restructuring from behavioral language data alone.

Subject of Research: A computational framework for predicting English-learning outcomes in secondary-school learner corpora

Article Title: Neuroplasticity-informed hierarchical and uncertainty-aware modeling of secondary-school english learning

Article References: Wang, H., Shaharom, M. S. N., Chin, H.-L., & He, Y. (2026). Neuroplasticity-informed hierarchical and uncertainty-aware modeling of secondary-school english learning. Frontiers in Molecular Neuroscience, 19, Article 1912397. https://doi.org/10.3389/fnmol.2026.1912397

Image Credits: AI Generated

DOI: 10.3389/fnmol.2026.1912397

Keywords: machine learning, English learning, uncertainty quantification, learner corpora, automated assessment, natural language processing, secondary education, predictive modeling, neuroplasticity, educational data, Neuroplasticity-informed, hierarchical